YOLO26: Optimized for AMD ROCm
YOLO26 is a real-time object detection model that detects and localizes axis-aligned bounding boxes across 80 COCO categories in a single forward pass. This repository packages inference for object detection using ONNX Runtime, exported and validated for AMD ROCm so it runs efficiently on AMD GPUs, CPUs, and NPUs.
This is based on the implementation of YOLO26 found here. This repository contains configurations and scripts optimized for AMD® ROCm™ platforms. You can use the yolo26_detect AMD scripts to reproduce results or export with custom configurations. More details on model performance can be found here.
Task Overview
Task: Object detection
Dataset: COCO val2017 (5,000 images, 80 categories)
Output metrics: mAP@0.5:0.95, mAP@0.5, mAP@0.75, AR@100, per-size mAP/AR (small/medium/large)
Model variants: Default is n. Override with
MODEL_SIZE=n/s/m/l/x.
GPU/NPU note: MIGraphX runtime-quantizes a single FP32 ONNX model to cover all GPU precisions. NPU execution requires BF16 — VitisAI quantizes the FP32 ONNX internally at runtime using
configs/vitisai/detect_bf16.json.
AMD ROCm Optimization
This model export has been adapted and validated for AMD Instinct™ / Radeon™ GPUs running ROCm, as well as AMD CPUs and AMD Ryzen AI NPUs. Key points:
- Validated backends: ONNX Runtime across CPU (FP32), GPU (MIGraphX Execution Provider — FP32/FP16/BF16/INT8), and NPU (VitisAI Execution Provider — BF16, auto-quantized internally).
- No code changes required versus the upstream Ultralytics YOLO26 implementation — only environment/runtime configuration differs.
- GPU first run: MIGraphX compiles the model on first use (30+ minutes); compiled kernels are cached and reused afterwards.
| Runtime | Precision | Backend | Hardware | Notes |
|---|---|---|---|---|
| ONNX Runtime | FP32 | CPU Execution Provider | AMD CPU | — |
| ONNX Runtime | FP32 / FP16 / BF16 / INT8 | MIGraphX Execution Provider | AMD Instinct™ / Radeon™ GPU (ROCm) | First run pays a 30+ minute graph-compilation cost |
| ONNX Runtime | BF16 | VitisAI Execution Provider | AMD Ryzen AI NPU | Auto-quantized internally |
Getting Started
For setup instructions, evaluation scripts, and custom configuration options, see the yolo26_detect on GitHub.
Model Details
Model Type: Object detection (single-pass CNN detector)
Base Model: YOLO26 (Ultralytics), size n default
Model Stats:
- Input:
(1, 3, 640, 640)float32 - Output:
(1, 300, 6)float32 - Precision tested: FP32 (CPU); FP32, FP16, BF16, INT8 (GPU); BF16 (NPU)
Accuracy Pipeline
COCO quality evaluation is fully implemented — make eval-<device> runs inference across all 5,000 COCO val2017 images and scores predictions with pycocotools, the reference implementation for axis-aligned boxes.
Higher mAP means the model's predicted boxes and classes agree more closely with ground truth across the dataset — 1.0 is perfect detection, 0.0 means no correct detections. In practice, values above ~0.5 for mAP@0.5:0.95 are considered strong for COCO-scale object detection.
Metrics Explained
| Metric | Description |
|---|---|
| mAP@0.5:0.95 | Primary COCO metric — mean AP averaged across IoU thresholds 0.5–0.95. The strictest, most holistic accuracy number: higher means boxes are both correctly classified and tightly localized across a range of overlap thresholds. |
| mAP@0.5 | AP at a single, looser IoU threshold of 0.5 (VOC-style) — a prediction only needs to overlap the ground-truth box by 50% to count as correct. Typically higher than mAP@0.5:0.95; reflects "did it find the object" more than "how precisely." |
| mAP@0.75 | AP at a stricter IoU threshold of 0.75 — the predicted box must overlap ground truth by 75%. Rewards precise localization, not just correct detection. |
| AR@100 | Average recall with up to 100 detections per image — of all ground-truth objects, what fraction did the model find. High AR means few missed objects; low AR means the model is overlooking objects regardless of precision. |
| mAP/AR-small/medium/large | mAP@0.5:0.95 and AR@100 broken down by object size. Small objects are typically the hardest; this breakdown exposes size-specific weaknesses a single aggregate score would hide. |
Accuracy Results
Full Dataset Evaluation (COCO val2017) — MODEL_SIZE=n:
| Device | Precision | mAP@0.5:0.95 | mAP@0.5 | mAP@0.75 | AR@100 |
|---|---|---|---|---|---|
| CPU | FP32 | 0.4024 | 0.5578 | 0.4356 | 0.6142 |
| GPU | FP32 | 0.4024 | 0.5578 | 0.4356 | 0.6142 |
| NPU | BF16 | 0.0020 | 0.0029 | 0.0020 | 0.0052 |
NPU accuracy anomaly: the NPU BF16 run above reports near-zero mAP despite running genuinely on-device (execution_provider: VitisAIExecutionProvider, device_fell_back_to_cpu: False — not a silent CPU fallback). This is well outside the normal BF16-quantization delta and looks like a broken NPU accuracy path (e.g. a postprocessing/decode mismatch specific to the NPU export) rather than expected precision loss, and needs investigation before treating this number as representative.
Dig Deeper
Want to explore the full evaluation scripts, config options, and other AMD-optimized model examples?
📂 View the full project on GitHub
The GitHub repository includes:
- Setup and prerequisites for ROCm environments
- Full COCO val2017 evaluation pipeline via
pycocotools - Per-operator latency profiling scripts (including NPU AI Analyzer integration)
- Benchmarking and reproduction instructions across CPU, GPU, and NPU